Visual exploration of machine learning results using data cube analysis

Abstract

As complex machine learning systems become more widely adopted, it becomes increasingly challenging for users to understand models or interpret the results generated from the models. We present our ongoing work on developing interactive and visual approaches for exploring and understanding machine learning results using data cube analysis. We propose MLCube, a data cube inspired framework that enables users to define instance subsets using feature conditions and computes aggregate statistics and evaluation metrics over the subsets. We also design MLCube Explorer, an interactive visualization tool for comparing models' performances over the subsets. Users can interactively specify operations, such as drilling down to specific instance subsets, to perform more in-depth exploration. Through a usage scenario, we demonstrate how MLCube Explorer works with a public advertisement click log data set, to help a user build new advertisement click prediction models that advance over an existing model.